fix: improve double-embedding fix with type safety and UPDATE path test (#3723)

Make isinstance checks numpy-safe by using `not isinstance(dict)` instead
of `isinstance(list)`, add type hints to existing_embeddings parameter,
and add regression test for the infer=True UPDATE path.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

mem0/memory/main.py

@@ -9,7 +9,7 @@ import uuid
 import warnings
 from copy import deepcopy
 from datetime import datetime, timezone
-from typing import Any, Dict, Optional
+from typing import Any, Dict, List, Optional, Union
 
 from pydantic import ValidationError
 
@@ -479,7 +479,8 @@ class Memory(MemoryBase):
 
                 msg_content = message_dict["content"]
                 msg_embeddings = self.embedding_model.embed(msg_content, "add")
-                mem_id = self._create_memory(msg_content, msg_embeddings, per_msg_meta)
+                # Pass embeddings as a dict so _create_memory can reuse the cached embedding
+                mem_id = self._create_memory(msg_content, {msg_content: msg_embeddings}, per_msg_meta)
 
                 returned_memories.append(
                     {
@@ -608,6 +609,9 @@ class Memory(MemoryBase):
 
                     event_type = resp.get("event")
                     if event_type == "ADD":
+                        # Ensure action_text has an embedding cached to avoid redundant API calls
+                        if action_text not in new_message_embeddings:
+                            new_message_embeddings[action_text] = self.embedding_model.embed(action_text, "add")
                         memory_id = self._create_memory(
                             data=action_text,
                             existing_embeddings=new_message_embeddings,
@@ -615,6 +619,9 @@ class Memory(MemoryBase):
                         )
                         returned_memories.append({"id": memory_id, "memory": action_text, "event": event_type})
                     elif event_type == "UPDATE":
+                        # Ensure action_text has an embedding cached to avoid redundant API calls
+                        if action_text not in new_message_embeddings:
+                            new_message_embeddings[action_text] = self.embedding_model.embed(action_text, "update")
                         self._update_memory(
                             memory_id=temp_uuid_mapping[resp.get("id")],
                             data=action_text,
@@ -1153,10 +1160,13 @@ class Memory(MemoryBase):
         capture_event("mem0.history", self, {"memory_id": memory_id, "sync_type": "sync"})
         return self.db.get_history(memory_id)
 
-    def _create_memory(self, data, existing_embeddings, metadata=None):
+    def _create_memory(self, data: str, existing_embeddings: Union[Dict[str, List[float]], List[float]], metadata=None):
         logger.debug(f"Creating memory with {data=}")
-        if data in existing_embeddings:
+        # existing_embeddings may be a dict (preferred) or a precomputed vector
+        if isinstance(existing_embeddings, dict) and data in existing_embeddings:
             embeddings = existing_embeddings[data]
+        elif not isinstance(existing_embeddings, dict):
+            embeddings = existing_embeddings
         else:
             embeddings = self.embedding_model.embed(data, memory_action="add")
         memory_id = str(uuid.uuid4())
@@ -1220,7 +1230,7 @@ class Memory(MemoryBase):
 
         return result
 
-    def _update_memory(self, memory_id, data, existing_embeddings, metadata=None):
+    def _update_memory(self, memory_id, data: str, existing_embeddings: Union[Dict[str, List[float]], List[float]], metadata=None):
         logger.info(f"Updating memory with {data=}")
 
         try:
@@ -1253,8 +1263,10 @@ class Memory(MemoryBase):
         if "role" not in new_metadata and "role" in existing_memory.payload:
             new_metadata["role"] = existing_memory.payload["role"]
 
-        if data in existing_embeddings:
+        if isinstance(existing_embeddings, dict) and data in existing_embeddings:
             embeddings = existing_embeddings[data]
+        elif not isinstance(existing_embeddings, dict):
+            embeddings = existing_embeddings
         else:
             embeddings = self.embedding_model.embed(data, "update")
 
@@ -1523,7 +1535,8 @@ class AsyncMemory(MemoryBase):
 
                 msg_content = message_dict["content"]
                 msg_embeddings = await asyncio.to_thread(self.embedding_model.embed, msg_content, "add")
-                mem_id = await self._create_memory(msg_content, msg_embeddings, per_msg_meta)
+                # Pass embeddings as a dict so _create_memory can reuse the cached embedding
+                mem_id = await self._create_memory(msg_content, {msg_content: msg_embeddings}, per_msg_meta)
 
                 returned_memories.append(
                     {
@@ -1651,6 +1664,11 @@ class AsyncMemory(MemoryBase):
                     event_type = resp.get("event")
 
                     if event_type == "ADD":
+                        # Ensure action_text has an embedding cached to avoid redundant API calls
+                        if action_text not in new_message_embeddings:
+                            new_message_embeddings[action_text] = await asyncio.to_thread(
+                                self.embedding_model.embed, action_text, "add"
+                            )
                         task = asyncio.create_task(
                             self._create_memory(
                                 data=action_text,
@@ -1660,6 +1678,11 @@ class AsyncMemory(MemoryBase):
                         )
                         memory_tasks.append((task, resp, "ADD", None))
                     elif event_type == "UPDATE":
+                        # Ensure action_text has an embedding cached to avoid redundant API calls
+                        if action_text not in new_message_embeddings:
+                            new_message_embeddings[action_text] = await asyncio.to_thread(
+                                self.embedding_model.embed, action_text, "update"
+                            )
                         task = asyncio.create_task(
                             self._update_memory(
                                 memory_id=temp_uuid_mapping[resp["id"]],
@@ -2229,10 +2252,13 @@ class AsyncMemory(MemoryBase):
         capture_event("mem0.history", self, {"memory_id": memory_id, "sync_type": "async"})
         return await asyncio.to_thread(self.db.get_history, memory_id)
 
-    async def _create_memory(self, data, existing_embeddings, metadata=None):
+    async def _create_memory(self, data: str, existing_embeddings: Union[Dict[str, List[float]], List[float]], metadata=None):
         logger.debug(f"Creating memory with {data=}")
-        if data in existing_embeddings:
+        # existing_embeddings may be a dict (preferred) or a precomputed vector
+        if isinstance(existing_embeddings, dict) and data in existing_embeddings:
             embeddings = existing_embeddings[data]
+        elif not isinstance(existing_embeddings, dict):
+            embeddings = existing_embeddings
         else:
             embeddings = await asyncio.to_thread(self.embedding_model.embed, data, memory_action="add")
 
@@ -2315,7 +2341,7 @@ class AsyncMemory(MemoryBase):
 
         return result
 
-    async def _update_memory(self, memory_id, data, existing_embeddings, metadata=None):
+    async def _update_memory(self, memory_id, data: str, existing_embeddings: Union[Dict[str, List[float]], List[float]], metadata=None):
         logger.info(f"Updating memory with {data=}")
 
         try:
@@ -2349,8 +2375,10 @@ class AsyncMemory(MemoryBase):
         if "role" not in new_metadata and "role" in existing_memory.payload:
             new_metadata["role"] = existing_memory.payload["role"]
 
-        if data in existing_embeddings:
+        if isinstance(existing_embeddings, dict) and data in existing_embeddings:
             embeddings = existing_embeddings[data]
+        elif not isinstance(existing_embeddings, dict):
+            embeddings = existing_embeddings
         else:
             embeddings = await asyncio.to_thread(self.embedding_model.embed, data, "update")
 

tests/test_memory.py

@@ -438,3 +438,137 @@ async def test_async_update_nonexistent_memory_raises_error(mock_sqlite, mock_ll
         await memory._update_memory("non-existent-id", "new data", {"new data": [0.1, 0.2]})
 
     mock_vector_store.update.assert_not_called()
+
+
+@patch('mem0.utils.factory.EmbedderFactory.create')
+@patch('mem0.utils.factory.VectorStoreFactory.create')
+@patch('mem0.utils.factory.LlmFactory.create')
+@patch('mem0.memory.storage.SQLiteManager')
+def test_add_infer_false_embeds_once(mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory):
+    """
+    Regression test for issue #3723: adding with infer=False should not trigger duplicate embedding calls.
+
+    Root cause: _create_memory expected a dict for existing_embeddings but received a raw list[float],
+    causing the cache check `data in existing_embeddings` to always fail and trigger a redundant embed.
+    """
+    embedder = MagicMock()
+    embedder.embed.return_value = [0.1, 0.2, 0.3]
+    embedder.config = MagicMock(embedding_dims=3)
+    mock_embedder_factory.return_value = embedder
+
+    mock_vector_store = MagicMock()
+    mock_vector_store.search.return_value = []
+    mock_vector_store.insert.return_value = None
+    mock_vector_store.get.return_value = None
+    telemetry_vector_store = MagicMock()
+    mock_vector_factory.side_effect = [mock_vector_store, telemetry_vector_store]
+
+    mock_llm_factory.return_value = MagicMock()
+    mock_sqlite.return_value = MagicMock()
+
+    from mem0.memory.main import Memory as MemoryClass
+    memory = MemoryClass(MemoryConfig())
+
+    memory.add("foo", user_id="test_user", infer=False)
+
+    assert embedder.embed.call_count == 1
+    mock_vector_store.insert.assert_called_once()
+
+
+@patch('mem0.utils.factory.EmbedderFactory.create')
+@patch('mem0.utils.factory.VectorStoreFactory.create')
+@patch('mem0.utils.factory.LlmFactory.create')
+@patch('mem0.memory.storage.SQLiteManager')
+def test_add_infer_true_caches_embedding_on_llm_rewrite(mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory):
+    """
+    Regression test for issue #3723 (infer=True path): when the LLM rewrites a fact during the
+    ADD action, the embedding should be computed once and cached, not computed again inside _create_memory.
+    """
+    embedder = MagicMock()
+    embedder.embed.return_value = [0.1, 0.2, 0.3]
+    embedder.config = MagicMock(embedding_dims=3)
+    mock_embedder_factory.return_value = embedder
+
+    mock_vector_store = MagicMock()
+    mock_vector_store.search.return_value = []
+    mock_vector_store.insert.return_value = None
+    mock_vector_store.get.return_value = None
+    telemetry_vector_store = MagicMock()
+    mock_vector_factory.side_effect = [mock_vector_store, telemetry_vector_store]
+
+    # LLM extracts fact "User likes Python", then ADD action rewrites to "The user enjoys Python"
+    mock_llm = MagicMock()
+    mock_llm.generate_response.side_effect = [
+        json.dumps({"facts": ["User likes Python"]}),
+        json.dumps({"memory": [{"id": "0", "text": "The user enjoys Python", "event": "ADD", "old_memory": None}]}),
+    ]
+    mock_llm_factory.return_value = mock_llm
+
+    mock_sqlite.return_value = MagicMock()
+
+    from mem0.memory.main import Memory as MemoryClass
+    memory = MemoryClass(MemoryConfig())
+
+    memory.add("I like Python", user_id="test_user", infer=True)
+
+    # embed should be called exactly twice:
+    # 1. For the extracted fact "User likes Python" (search)
+    # 2. For the rewritten text "The user enjoys Python" (pre-cached before _create_memory)
+    # It should NOT be called a 3rd time inside _create_memory
+    assert embedder.embed.call_count == 2
+    mock_vector_store.insert.assert_called_once()
+
+
+@patch('mem0.utils.factory.EmbedderFactory.create')
+@patch('mem0.utils.factory.VectorStoreFactory.create')
+@patch('mem0.utils.factory.LlmFactory.create')
+@patch('mem0.memory.storage.SQLiteManager')
+def test_update_infer_true_caches_embedding_on_llm_rewrite(mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory):
+    """
+    Regression test for issue #3723 (infer=True UPDATE path): when the LLM rewrites a fact during
+    an UPDATE action, the embedding should be computed once and cached, not computed again inside _update_memory.
+    """
+    embedder = MagicMock()
+    embedder.embed.return_value = [0.1, 0.2, 0.3]
+    embedder.config = MagicMock(embedding_dims=3)
+    mock_embedder_factory.return_value = embedder
+
+    # Existing memory that will be matched for update
+    existing_memory = MockVectorMemory(
+        memory_id="existing-mem-id",
+        payload={
+            "data": "User likes Python",
+            "hash": "abc123",
+            "created_at": "2025-01-01T00:00:00+00:00",
+        },
+    )
+
+    mock_vector_store = MagicMock()
+    mock_vector_store.search.return_value = [existing_memory]
+    mock_vector_store.get.return_value = existing_memory
+    mock_vector_store.insert.return_value = None
+    mock_vector_store.update.return_value = None
+    telemetry_vector_store = MagicMock()
+    mock_vector_factory.side_effect = [mock_vector_store, telemetry_vector_store]
+
+    # LLM extracts fact "User loves Python now", then UPDATE action rewrites to "The user loves Python"
+    mock_llm = MagicMock()
+    mock_llm.generate_response.side_effect = [
+        json.dumps({"facts": ["User loves Python now"]}),
+        json.dumps({"memory": [{"id": "0", "text": "The user loves Python", "event": "UPDATE", "old_memory": "User likes Python"}]}),
+    ]
+    mock_llm_factory.return_value = mock_llm
+
+    mock_sqlite.return_value = MagicMock()
+
+    from mem0.memory.main import Memory as MemoryClass
+    memory = MemoryClass(MemoryConfig())
+
+    memory.add("I love Python now", user_id="test_user", infer=True)
+
+    # embed should be called exactly twice:
+    # 1. For the extracted fact "User loves Python now" (search)
+    # 2. For the rewritten text "The user loves Python" (pre-cached before _update_memory)
+    # It should NOT be called a 3rd time inside _update_memory
+    assert embedder.embed.call_count == 2
+    mock_vector_store.update.assert_called_once()